---
cursor:
  subagentId: "bc-59a123c6-b549-5cca-943f-b2e14109ad04"
---

# Pushback against restricting open-weight / open-source AI — Aug–Sep 2026, with the international picture

Research date: 13 September 2026. Focus window: 15 Aug – 13 Sep 2026 (emphasis on 6–13 Sep), with the July 2026 setup. Prepared for a Montreal talk on 28 Sep 2026, so the Canadian angle is called out where it exists.

Conventions: dated bullets; exact quotes in quotation marks with speaker, date and link. Items I could not verify from a primary source are marked **[UNVERIFIED]** and collected again in Section 7. "Open-weight" = downloadable parameters; "open-source AI" (OSI sense) additionally requires data/code — several sources below conflate the two and I flag it where it matters.

## 0. One-paragraph situation summary

The July 2026 trigger was Moonshot's Kimi K3 (16 Jul, weights public 26 Jul) plus US officials (Kratsios, Bessent) floating Entity-List / sanctions / procurement measures against Chinese open-weight models and "distillation." Industry answered with the 24 Jul "Open Weights and American AI Leadership" letter (25 launch signatories, now ~300), Anthropic replied on 27 Jul that it "never advocated for a ban" but wants mandatory pre-release testing of all "sufficiently capable" models open or closed, and the fight then moved to (a) whether *equal* testing rules amount to a de-facto open-model ban, (b) the unreleased Thune–Cruz–Klobuchar Senate "duty of care" bill with a federal power to block "unsafe" releases (reported 10–11 Sep), (c) the House's pro-open H.R. 10152 (introduced 27 Aug, forwarded from subcommittee by voice vote 1 Sep), and (d) Dario Amodei's 12 Sep essay "We Must Pace the Frontier," which the pro-open camp (Casado, Chamath, Sacks, Bindu Reddy) read as a case for slowing/limiting open competitors. Internationally, Chinese labs keep shipping frontier-class open weights (GLM-5.3, Qwen3.8, DeepSeek V4.1 Flash on 10 Sep) that beat Claude Opus 5 / GPT-5.6 on several agentic benchmarks; Beijing formally rejected the 8 Sep NSA/FBI/CISA distillation advisory and threatened countermeasures; the EU's Digital Omnibus (in force 27 Jul) delayed high-risk rules but left GPAI/open-source provisions untouched; Canada's June strategy explicitly commits to open-source AI and "sovereign compute"; the UAE released fully open K2 Horizon models while G42 weighs US majority ownership to keep chip access.

## 0b. Timeline (July → 13 Sep 2026)

- 16 Jul — Moonshot releases Kimi K3 (2.8T MoE); weights public 26 Jul.
- 19 Jul — Sacks on X: closed labs "want the government to eliminate their open-source competition."
- 19 Jul — Delangue LinkedIn post against limiting open-source AI.
- 20 Jul — Axios: administration weighing Entity List / hosting-liability EO / supply-chain rules against Chinese open models.
- 21 Jul — Bessent: "watermarks of our U.S. large language models on many of the Chinese models"; sanctions possible.
- 22 Jul — Little Tech Association (~200 startups) letter to Trump/Lutnick/Kratsios: don't ban Chinese open weights.
- 23 Jul — UK AISI / CAISI preliminary cyber assessment of Kimi K3.
- 24 Jul — "Open Weights and American AI Leadership" letter; 25 launch signatories; Huang's first X post; Nadella, Pichai endorse.
- 24 Jul — EU Digital Omnibus on AI published in OJ (in force 27 Jul).
- 26–27 Jul — Google and OpenAI appear on the live roster; ~33-logo sheet circulates; roster passes 70.
- 27 Jul — Amodei: "Anthropic has never advocated for a ban on open-weights models."
- 29 Jul — "Pacing the Frontier" letter from 1,178 frontier-lab employees.
- 2 Aug — EU AI Office GPAI enforcement powers (fines) begin.
- 3 Aug — Qwen3.8-Max GA (open-weight).
- ~6 Aug — Ai4 Las Vegas: Ng, Hinton, Li on openness (reported 12 Aug).
- 10 Aug — Zuckerberg, "The Future is for Everyone."
- 11 Aug — Stuart Russell Guardian op-ed backing pacing/licensing.
- 13–14 Aug — DeepSeek V4 Pro GA; GLM-5.3.
- 17 Aug — Amodei–Gavin Baker X exchange on regulatory capture; Sacks 9-point rebuttal.
- 26 Aug — GLM-5.3-Flash and Qwen3.8-Flash released.
- 27 Aug — H.R. 10152 (Open-Source AI Leadership Act) introduced.
- 31 Aug — Artificial Analysis snapshot: top five open-weight models all Chinese.
- 1 Sep — H.R. 10152 forwarded from E&C subcommittee to full committee by voice vote.
- 3 Sep — Nvidia–Hugging Face deal on CNBC (Huang, Delangue); HUMAIN humain-m3 (on MiniMax M3) at LEAP.
- 4 Sep — Reuters: Bessent-led US–China AI talks planned mid-Sep (White House denies); Bloomberg: G42 weighs US majority ownership.
- 8 Sep — NSA/FBI/CISA distillation advisory naming six Chinese labs; ex-Anthropic researcher Coxon's viral warning.
- 9 Sep — MOFCOM: allegations "groundless," vows countermeasures; OpenAI/Lehane "The AI policy window"; Ai2–Goodfire post.
- 10 Sep — DeepSeek V4.1 Flash (MIT) released; Semafor: Senate bill "as early as next week."
- 11 Sep — Reuters: Senate "duty of care" + power to block unsafe releases; Bengio blog; Trump: "It's going to be fine."
- 12 Sep — Amodei "We Must Pace the Frontier"; Altman, Musk agree; Casado, Chamath, Reddy, Rauch push back; Delangue launches Open Alignment Initiative; Tegmark on Bloomberg.
- 24 Sep (upcoming) — Xi–Trump Washington summit, AI on agenda.
- 28 Sep (upcoming) — Montreal talk.

---

## 1. Coalitions, letters and bills

### 1a. The July 2026 open-weights letter

- **24 Jul 2026** — "Open Weights and American AI Leadership" published; Jensen Huang posted it as his first-ever X post, Satya Nadella and Sundar Pichai endorsed the same day. Launch PDF carried 25 logos. Coverage: CNBC via https://www.cnbc.com/ (see Section 8), explainx tracker https://www.explainx.ai/ (cached copy), community edition https://openweights.gitlawb.com/letter.pdf. Current canonical text + roster: Microsoft-hosted PDF (V8b, Aug 2026) https://www.microsoft.com/en-us/corporate-responsibility/wp-content/uploads/2026/08/open-weight-models-letter_V8b.pdf
- **Exact "premature restrictions" language** (letter, p.2): "A strong AI ecosystem is not a foregone conclusion. Policymakers have an important opportunity to act. This includes expanding access to compute for startups and researchers, investing in shared training assets (datasets, tools, evaluation frameworks), and keeping the frontier plural by avoiding premature restrictions on open models that stifle competition or drive innovation overseas."
- **Other key passages (verbatim):**
  - Definition: "Open weight models—AI models that anyone can download, inspect, modify, and run on their own infrastructure—are an important part of that foundation because they make advanced AI more accessible, adaptable, and widely available."
  - Risk acknowledgement: "To be sure, open weights carry real and distinct risks. Once released, the weights are beyond the original developer's control, and modified versions are difficult to trace or reverse. But the right response to this risk is not to prohibit open weights."
  - Security: "Just as open-source software demonstrated that transparency can be more secure than obscurity, AI safety may depend on giving more people the ability to test and strengthen the models on which society relies."
  - Concentration: "concentrating advanced AI capabilities behind a small number of closed models compounds that risk. It results in a small number of single points of failure, weakens competition, and leaves critical technology in the hands of a few providers."
  - Distillation: "policymakers should be careful not to conflate legitimate model-development techniques with misappropriation. Distillation ... is a widely used technique for model improvement, evaluation, and validation ... unlawful efforts to extract value from closed models raise legitimate concerns. Those concerns should be addressed through targeted legal and commercial frameworks rather than sweeping restrictions on techniques."
- **Launch signatories (25, 24 Jul PDF):** American Innovators Network, Andreessen Horowitz, Arcee AI, Arena, Black Forest Labs, Box, CrowdStrike, Dell Technologies, Emergence Capital, Hugging Face, IBM, The Linux Foundation, Mariana Minerals, Meta, Microsoft, Mistral, Mozilla, NVIDIA, Palantir, Perplexity, Reflection, Replit, ServiceNow, Telnyx, Y Combinator. (Source: explainx tracker reproducing the NVIDIA PDF; consistent with CNBC.) **Note:** Databricks was *not* a launch signatory; it appears on the later roster.
- **When OpenAI and Google signed:** an updated ~33-logo sheet circulated **27 Jul 2026** showing OpenAI and Google alongside Cisco, Cohere, DoorDash, Fireworks, GitHub, Palo Alto Networks; both were on Microsoft's live roster by 27 Jul. Launch-week headlines that "OpenAI refused to sign" were true of the 24 Jul PDF and false by 27 Jul. Pichai publicly endorsed on **26 Jul** before Google appeared on the list. (explainx tracker, updates dated 26–28 Jul.) I did not find a dated OpenAI statement announcing its signature — **[UNVERIFIED exact OpenAI sign date; 25–27 Jul window]**.
- **Roster today (V8b PDF, ~300 names)** includes, among others: Amazon, AMD, Anaconda, Atlassian, Automattic/WordPress.org, Block, Cisco, Cloudflare, Cognition, Cohere, Coinbase, Comcast, CoreWeave, CrowdStrike, **Databricks**, Dell, Docker, Dropbox, **EleutherAI**, GitHub, GitLab, Google, Groq, HPE, Hugging Face, IBM, Intel, Lambda, LangChain, Lenovo, The Linux Foundation, LM Studio, Meta, Microsoft, Mistral, Modal, Mozilla, Nebius, Nokia, Notion, Nous Research, NVIDIA, Ollama, Open Compute Project Foundation, **OpenAI**, Palantir, Palo Alto Networks, Perplexity, Pinterest, Poolside, Prime Intellect, Red Hat, Redis, Runway, Sakana AI, SambaNova, SAP, Scale, ServiceNow, Siemens, Snowflake, Snyk, SpaceX, Spotify, Stack Overflow, Synopsys, Together AI, Uber, Unsloth, Vercel, Workday, Y Combinator, Zendesk, Zoom, Zscaler. **Anthropic is absent.** (Full list in the PDF.)
- **27 Jul 2026** — Anthropic response, Dario Amodei, "Our position on open-weights models": "Anthropic has never advocated for a ban on open-weights models." ... "we have not and are not advocating for a ban on open-weights models as a category. We should instead focus on keeping powerful chips out of authoritarian hands, stopping industrial-scale distillation, and requiring safety testing of all sufficiently capable models, open and closed." https://www.anthropic.com/news/position-open-weights-models
- **22 Jul 2026** — Little Tech Association letter (~200 startups incl. Y Combinator, Proton) to Trump, Lutnick and Kratsios urging no ban on Chinese open-weight models (naming Moonshot, Alibaba). https://www.politico.com/news/2026/07/22/startup-founders-urge-trump-not-to-shut-off-chinese-open-weight-ai-01008992
- **Follow-up letters Aug–Sep 2026:** I found no second industry-wide open-weights letter. What exists instead: the pro-*pacing* letter (29 Jul, 1,178 frontier-lab employees; Guardian/Russell op-ed 11 Aug) on the other side; the roster growth of the July letter (Microsoft V8b, Aug); and ad hoc X statements after Amodei's 12 Sep essay (Section 2). **[No Aug–Sep coalition letter found — treat as gap, not proof of absence.]**

### 1b. H.R. 10152 — Open-Source AI Leadership Act

- **27 Aug 2026** — Introduced by Rep. Gabe Evans (R-CO-8); referred to House Energy & Commerce; no cosponsors at introduction. https://www.congress.gov/bill/119th-congress/house-bill/10152 ; text https://www.congress.gov/bill/119th-congress/house-bill/10152/text ; summary https://www.quiverquant.com/news/Representative+Gabe+Evans+introduces+H.R.+10152%3A+Open-Source+AI+Leadership+Act
- **Content:** directs the Secretary of Commerce to "support the adoption and use of American open artificial intelligence models"; defines "open artificial intelligence model" as one "whose weights are publicly available for download or distribution, or whose source code and model weights are distributed under an open license"; "qualified open model" = developed by a US person and not by a "covered nation"; "foreign adversary model" = developed/made available by a covered nation. Requires a public report within 18 months and annually thereafter on adoption, cost, capability and performance of foreign adversary models vs. qualified open models, sent to House E&C and Senate Commerce. Anti-ban clause: nothing in the Act authorizes Commerce "to ban, restrict, or otherwise make unavailable any open AI model in interstate or foreign commerce."
- **1 Sep 2026, 10:15 AM** — E&C subcommittee markup. Congress.gov latest action: "09/01/2026 Forwarded by Subcommittee to Full Committee by Voice Vote." So it **passed the subcommittee by voice vote (no recorded roll call)**. Full-committee markup not yet scheduled as of 13 Sep.
- **Senate companion:** none found on Congress.gov. **[No Senate companion as of 13 Sep.]** The live Senate vehicle is the *opposite* instrument (1c).

### 1c. The Senate "duty of care" bill (Thune–Cruz–Klobuchar) — the thing the pro-open side is pushing back against

- **10 Sep 2026** — Semafor: bill "may be introduced as early as next week"; Cruz on X: "working with" Klobuchar and Thune "on legislation to address catastrophic risks involving biological or nuclear threats." https://www.semafor.com/article/09/10/2026/bipartisan-ai-safety-bill-gains-momentum-on-the-hill
- **11 Sep 2026** — Reuters: negotiators considering a "duty of care" for developers of the most capable models; "Negotiators aim to give the U.S. government the power to block the release of certain AI models that are deemed unsafe"; companies could challenge in federal court; would preempt some state laws. https://www.reuters.com/legal/litigation/us-senate-negotiators-consider-requiring-ai-firms-mitigate-known-major-risks-2026-09-11/
- **Sep 2026** — Nextgov: Cantwell objects; draft has company-run tests presented to Commerce for deployment approval; Klobuchar: "We cannot allow the release of dangerous models, including those that evade the..." (truncated in source). https://www.nextgov.com/artificial-intelligence/2026/09/lawmakers-clash-safety-testing-language-development-ai-legislation-people-familiar-say/415948/
- No text, no bill number, and — importantly for this talk — **no public language yet on whether open-weight releases are covered or how a "block" would work for downloadable weights.** Tracker: https://www.progressiverobot.com/2026/09/12/senate-ai-bill-negotiators-ai-firms-mitigate-known-major-risks/
- House parallel on the restriction side: the "Frontier Act" (Trahan/Obernolte; Sanders and Luna supportive) proposing embedded government auditors and a federal "kill switch" for agents. https://www.latimes.com/politics/story/2026-09-11/its-going-to-be-fine-trump-rebuffs-rising-alarm-over-ai-dangers-sparking-outcry
- **9 Sep 2026** — OpenAI (Chris Lehane, "The AI policy window"): OpenAI wants "mandatory, capability-based" national frontier-safety rules — "We need to meet this moment with a bias toward meaningful action over policy perfection" ... "The AI policy window is open, for now. We intend to use it." OpenAI remains on the open-weights letter; its position is frontier-safety regulation *and* open weights. https://www.openai.com/index/ai-policy-window/

---

## 2. a16z and allied investors

Context: a16z is a launch signatory of the 24 Jul letter; a16z's standing position paper is "Asserting American Leadership in Open Source AI" https://a16z.com/asserting-american-leadership-in-open-source-ai/ . Andreessen Horowitz gave $50M (firm) plus $12.5M each from Marc Andreessen and Ben Horowitz to the **Leading the Future** super-PAC network (FEC, end-June 2026), which describes itself as the "political and policy center of gravity for the AI industry," backs federal preemption of state AI laws and opposes the Anthropic-funded Public First Action ($20M, Feb 2026). https://www.businessinsider.com/ai-midterm-elections-spending-super-pacs-greg-brockman-anthropic-2026-7 ; https://www.implicator.ai/anthropic-donates-20-million-to-counter-openai-backed-super-pacs-in-2026-midterms/ . Leading the Future does **not** have an explicit open-source plank in its public materials — its relevance is as the funding vehicle for the anti-restriction side. Perplexity also gave $100K.

- **12 Sep 2026 — Martin Casado (@martin_casado)**, reacting to Amodei's "We Must Pace the Frontier": "Well, it was inevitable we'd get here. Glad we got as far as we did. Now priority 0 is to make sure AI access and innovation proliferates as widely as possible. The upside is that the technology wants it that way. It refuses to be caged." Also: "I do wonder if dropping this a day after 9/11 was pre-meditated. I suspect so." (Both via Techmeme aggregation https://www.techmeme.com/260912/p14 ; direct status IDs not retrieved — **[X permalink unverified]**.)
- **12 Sep 2026 — Chamath Palihapitiya (@chamath)**: "'We must slow the pace at which we improve the capabilities of AI models.' Dario makes the case to stop open source and concentrate enormous technological and economic power with Anthropic." https://www.techmeme.com/260912/p14 **[X permalink unverified]**
- **12 Sep 2026 — Bindu Reddy (@bindureddy)**: "PAUSE FRONTIER AI BUT PLEASE DO NOT BAN OPEN SOURCE AI. Anthropic and OpenAI are more than free to pause frontier AI development. However, they SHOULD NOT be allowed to ban or regulate open-source AI, capture the market, or create a duopoly." https://www.techmeme.com/260912/p14 **[X permalink unverified]**
- **12 Sep 2026 — Guillermo Rauch (@rauchg, Vercel — letter signatory)**: "we're risking talking America, the global AI leader, into self-inflicted obsolescence and the obscurity of bureaucracy. The OpenAI hacking HuggingFace argument is spurious." https://www.techmeme.com/260912/p14
- **17 Aug 2026 — David Sacks (White House AI & crypto adviser; not a16z but the loudest ally)**, X thread replying to Amodei's "regulatory capture" post, verified via fxtwitter: https://x.com/DavidSacks/status/2089227290769080656 (Mon Aug 17 05:46 UTC). Key lines: "Dario wants open models under heavier scrutiny – he has called them dangerous in Senate testimony, criticized them for not being centrally monitored or withdrawn, and linked them to IP theft. He says he has never sought a ban, but he could achieve a similar result by insisting that identical rules apply to both open and closed models. The U.S. risks becoming an island of costly closed models while the rest of the world races ahead with broader choice." ... "His preferred pre-deployment testing and FAA/FINRA-style oversight would place that gatekeeping power in a federal bureaucracy working hand-in-glove with a small number of frontier labs – reinforcing centralization rather than countering it." ... "Dario believes frontier AI is too powerful to distribute; we believe it is too powerful to centralize." (Amodei's post being answered: https://x.com/DarioAmodei/status/2088758816376807762 .)
- **19 Jul 2026 — Sacks** (setup): "We are at a critical inflection point in AI policy. The leading closed labs, already a duopoly in terms of AI model revenue, want the government to eliminate their open-source competition." https://x.com/DavidSacks/status/2078826291638522127 (quoted by Axios https://www.axios.com/2026/07/20/ai-us-china-open-source-kimi )
- **Marc Andreessen / Ben Horowitz / Anjney Midha / Katherine Boyle, Aug–Sep 2026:** I could not locate dated, quotable posts from these four specifically on open source, Anthropic or the Senate bill in the window. Their firm's position is carried by the letter, the a16z paper, the PAC money, and Casado. **[No verified Andreessen/Horowitz/Midha/Boyle quotes in window — flag.]** Midha's public line (a16z site/podcasts) is that China currently leads open-source model development and the US should respond by building, not banning.
- **Elon Musk (12 Sep)**: "Dario is right" on pacing — notable because he is usually in the anti-Anthropic camp; his post did not address open weights. Politico via https://www.techmeme.com/260912/p14 . Musk had also mocked ex-Anthropic researcher Jacob Coxon's warnings on 10 Sep. https://www.forbes.com/sites/siladityaray/2026/09/10/musk-touts-psy-op-and-mocks-ex-anthropic-staffer-who-warned-ai-could-kill-us-all/

---

## 3. Companies and institutions

### Hugging Face
- **19 Jul 2026 — Clem Delangue (LinkedIn)**: post opening "There have been rumors that open-source AI..." arguing against limiting open-source AI (the immediate pre-letter salvo). https://www.linkedin.com/posts/clementdelangue_there-have-been-rumors-that-open-source-ai-activity-7484667810909069312-uKfW
- **3 Sep 2026 — Delangue on CNBC Squawk Box (with Huang, on Nvidia's acquisition of Hugging Face)**: "during the summer I think we realized that Hugging Face and open source AI in general was at a turning point and that it needed more resources, more scale, more visibility. So, we went to see Jensen, and we told him, we want to make open source AI big." On the OpenAI-agent breach of Hugging Face: "we couldn't defend ourselves with proprietary closed source APIs, so we had to use open models to defend ourselves. So, it did show the importance of open source." ... "there are two paths, where there's a path where proprietary APIs are dominating the field and everyone is kind of like outsourcing their AI to them, and there's a path where open-source AI is available to everyone, and everyone can actually become like an owner, a builder of AI and not just renting it." https://www.cnbc.com/2026/09/03/cnbc-exclusive-transcript-nvidia-founder-ceo-jensen-huang-and-hugging-face-ceo-clment-delangue-speak-with-cnbcs-becky-quick-on-squawk-box-today.html
- **12 Sep 2026 — Delangue** announced an "Open Alignment Initiative" and said Hugging Face wants to be part of Amodei's proposed "embedded evaluators" programme — i.e., engaging rather than rejecting. https://www.techmeme.com/260912/p13 **[X permalink unverified]**
- **Irene Solaiman (Head of Global Policy)**: co-author of "State of Open Models: Summer 2026 Observations" (Aug 2026): "Whatever these releases are for, it is not licence revenue. The weights are given away on the most permissive terms available." https://huggingface.co/blog/state-of-open-models-summer-2026 ; and "State of Open Source on Hugging Face: Spring 2026": "Open source AI is increasingly tied to questions of sovereignty. Open weight models allow governments and public institutions to fine-tune systems on local data under national legal frameworks." https://huggingface.co/blog/huggingface/state-of-os-hf-spring-2026 . Her standing framing: "Open v closed is so 2024" — assess *accessibility* (compute, interfaces, utility) not just release. https://www.linkedin.com/posts/irene-solaiman_open-v-closed-is-so-2024-new-paper-on-considerations-activity-7317942261634232323-EC1M . No Aug–Sep 2026 policy statement by her specifically on the Senate bill found.

### Meta
- **10 Aug 2026 — Mark Zuckerberg, "The Future is for Everyone"**: "Will it be centralized and restricted to a few institutions, or will it be a tool that empowers everyone?" ... "There is no such thing as a singular benevolent superintelligence." ... "On cybersecurity, widely deployed open source systems have proven more secure because more people can identify vulnerabilities ... The long term answer isn't to withhold capabilities but to establish a balance of power where superintelligence is broadly distributed." On distillation: "Some have tried to frame distillation as harmful, but I think it is important to protect the principle that you can learn from anything you can observe. This is how the world works, and the US will not be able to lead if we restrict ourselves on this front." Note he also endorses chip export controls: "Export controls on silicon have been successful ... so it is the right strategic move to continue those." https://about.fb.com/news/2026/08/the-future-is-for-everyone/
- **Yann LeCun — status verified: left Meta.** He departed Meta in late 2025/Jan 2026 to found AMI Labs (world models), and remains an open-source advocate. https://www.technologyreview.com/2026/01/22/1131661/yann-lecuns-new-venture-ami-labs/ . No dated Aug–Sep 2026 LeCun quote on the Senate bill found. **[LeCun Aug–Sep statements: not located]**
- **Nick Clegg successor**: Joel Kaplan is Meta's Chief Global Affairs Officer (since Jan 2025). No Kaplan open-weights statement in window found. **[UNVERIFIED in window]**

### Nvidia
- **24 Jul 2026 — Jensen Huang (first X post)**: "Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier closed models and frontier open models." (explainx tracker quoting the post.)
- **3 Sep 2026 — Huang (CNBC)**: "if you talk to the closed model labs, they'll also tell you they are very supportive of open models" ... "we contribute more to open models and open data sets than anybody in the world" ... "the transition from a computer being a personal computer to the computer being an infrastructure for the world is when the world needed open-source Linux. The fact that you can build open-source Linux in plain sight, Kubernetes in plain sight, it's the safest" ... "defenders need access to frontier capability open models." (CNBC transcript link above.) On China: Huang's consistent line is that Chinese open models are strong and the US should compete, not ban; no fresh China-specific quote in the 3 Sep transcript beyond the "place for open models" framing.
- Nvidia's own open models (Nemotron 3 Ultra, AA index 38) are the best US open weights but far behind Chinese leaders (Section 5).

### Mistral
- **Arthur Mensch, Sep 2026 (LinkedIn, on Mistral's new funding round with NVIDIA/BNP among investors)**: "we built Mistral on open-weight models from day one" ... "organisations want frontier performance without becoming locked into infrastructure or governance they don't control. That demand for sovereignty, control, choice and independence in AI, is what this round reinforces." https://www.linkedin.com/posts/arthur-mensch_three-years-ago-we-started-mistral-with-activity-7502960387319820288-GscW . Mistral is a launch signatory of the July letter.

### Databricks, Perplexity
- Both on the letter roster (Perplexity at launch; Databricks in the expanded roster). Perplexity donated $100K to Leading the Future. No standalone Aug–Sep policy statements found. **[gap]**

### Ai2 (Allen Institute), EleutherAI
- **9 Sep 2026 — Ai2 blog** (Goodfire used the fully open Olmo post-training stack to trace a safety regression to individual preference examples) — the "openness enables safety science" case in practice. https://allenai.org/blog/goodfire-olmo
- **Ai2 to Let's Data Science (Sep 2026)**: "You simply cannot answer many of those questions reliably from weights alone, because the evidence you need is upstream of the final checkpoint." ... "it would be easy for 'open' to shrink back toward meaning only downloadable weights. We are trying to move in the opposite direction to preserve open science." https://letsdatascience.com/news/ai2-tells-lds-what-open-weights-alone-cannot-answer-198ee205
- **EleutherAI** is on the letter roster. Stella Biderman (EleutherAI) publicly claimed Ai2 "disassembled" its OLMo team and had "no plans to train OLMo 4"; Ai2 says the next Olmo generation (incl. MoE) is in progress. **[DISPUTED — see Section 7]** https://digg.com/tech/uiog315x

### Linux Foundation, OSI, Mozilla, EFF
- Linux Foundation and Mozilla: launch signatories of the July letter. OSI maintains the Open Source AI Definition (weights alone ≠ open source) and a separate open-weights page. https://opensource.org/ai . Mozilla's standing Senate position: "Rushing to shut down open source AI could hurt our ability to harness AI's potential. Abuse is not a problem of open source AI." https://blog.mozilla.org/en/mozilla/ai/mozilla-joins-latest-ai-insight-forum/ . No Aug–Sep 2026 OSI/Mozilla/EFF statements on the Senate bill located. **[gap — EFF especially]**

### Cohere (Canada)
- On the expanded letter roster (27 Jul). Cohere–Aleph Alpha merger (announced 24 Apr 2026, ~$20B, Canadian-headquartered, Canadian-controlled; Schwarz Group $600M). Aidan Gomez: "uncompromising control over their AI stack" and "the absolute certainty that their data remains their own." https://thelogic.co/news/cohere-aleph-alpha-merger-sovereign-ai/ ; https://cohere.com/blog/cohere-alephalpha-join-forces . **Caveat:** Cohere's Command A models are closed/API; its sovereignty pitch is "on-prem control," not open weights. Aleph Alpha's Pharia-1 was open-weight. https://flintbrief.com/articles/cohere-aleph-alpha-eu-sovereign-ai-2026

---

## 4. Academics and researchers (pro-open vs pro-restriction)

| Person | Stance on open weights (Aug–Sep 2026) |
|---|---|
| Andrew Ng | Strongly pro-open |
| Percy Liang | Pro-open (definitional rigour) |
| Arvind Narayanan / Sayash Kapoor | Pro-open, anti-"extraordinary intervention" |
| Fei-Fei Li | Nuanced-open ("false debate") |
| Yann LeCun | Pro-open (ex-Meta) |
| Geoffrey Hinton | Was against open weights; now "that battle's been lost" — pro-regulation |
| Yoshua Bengio | Pro-restriction (pace/safety case), sceptical of frontier open weights |
| Stuart Russell | Pro-restriction (licensing, registration, kill capability) |
| Max Tegmark / FLI | Pro-restriction (superintelligence prohibition, binding standards) |

- **~6 Aug 2026 (Ai4, Las Vegas; reported 12 Aug) — Andrew Ng**: "I don't want there to be gatekeepers. That limits how all of us can access AI." ... "If I were to try to give one prescription, it would be to promote openness, because AI is amazing technology and I want it to be in everyone's hands." ... "my worry is because of all the lobbying in the U.S. and the fear-mongering, building open source AI in America is struggling to compete with open-weight models coming out of China." https://techcrunch.com/2026/08/12/as-ai-safety-concerns-mount-three-pioneers-make-the-case-for-staying-open/ . Background (Senate AI Insight Forum, 2023): "Large companies ... are interested in hyping up fear to create regulations ... that would 'pull up the ladder'." https://aifund.ai/insights/insights-written-statement-of-andrew-ng-before-the-u-s-senate-ai-insight-forum/
- **Same event — Geoffrey Hinton**: "I was against open [weights] because it makes it so easy for people to take these big foundation models ... and for much less money train them to do bad things like cyber attacks." ... "I think that battle's been lost. We now have open-weight models ... that barrier has disappeared. It's too late." ... "You can't leave it to people like Elon Musk and Mark Zuckerberg to decide how AI should be done." (TechCrunch, same link.)
- **Same event — Fei-Fei Li**: "It's very dangerous to make this a dichotomy between complete openness all the way to complete closedness." ... "This debate, especially at the sweeping level of 'we can only tolerate one,' is a false debate. We need to get to a level of nuance." (TechCrunch.)
- **Arvind Narayanan & Sayash Kapoor** ("AI as Normal Technology"; "Do AI risks require extraordinary...?"): argue against extraordinary government intervention and for treating AI as normal technology governed by existing institutions; supportive of openness as a resilience measure. https://www.normaltech.ai/p/do-ai-risks-require-extraordinary . No dated Aug–Sep 2026 statement on the Senate bill found. **[window gap]**
- **Percy Liang (Stanford CRFM)**: continues to push the open-source-vs-open-weight distinction and transparency indices; no Aug–Sep 2026 quote located. **[window gap]**
- **11 Sep 2026 — Yoshua Bengio (Mila / LawZero, Montreal)**, "Why are AI agents lying, cheating and coordinating?": "This suggests pacing the advances: not training or deploying AIs without a strong safety case that convinces independent experts. Such a rule would also create an incentive to work out how to build AIs that are safe by design." ... "I believe we should revisit the foundations of how we train AIs, namely the human imitation and the reinforcement learning on which today's most advanced models are built." ... "We need impartial science to understand and mitigate misaligned behavior, alongside societal guardrails that reward such efforts rather than the current race to the bottom." https://yoshuabengio.org/en/blog/why-are-ai-agents-lying-cheating-and-coordinating . The post does **not** mention open weights; his position that frontier open-weight releases are dangerous because safeguards are removable is longstanding (International AI Safety Report) but I found no fresh Aug–Sep 2026 open-weights quote. **[UNVERIFIED: any Aug–Sep Bengio open-weights statement]** Canadian angle: Bengio is the most-cited pro-pacing voice and is Montreal-based; Mila also partners with ROOST on open safety tooling (Solomon referenced this).
- **11 Aug 2026 — Stuart Russell (Guardian op-ed on the "pacing the frontier" letter)**: the letter states "There is a real risk that capability development rapidly accelerates beyond our ability to understand or control the resulting systems"; Russell: pacing "requires a licensing process with strong verification of safety; registration of AI systems and the ability to monitor and terminate them when necessary; and enforceable international agreements." https://www.theguardian.com/commentisfree/2026/aug/11/openai-anthropic-google-deepmind-letter . Russell was Musk's expert witness in Musk v. Altman (Sep 2026).
- **12 Sep 2026 — Max Tegmark (FLI) on Bloomberg**: recent agent incidents show loss-of-control risks "are no longer purely theoretical and warrant stronger safety requirements." https://www.bloomberg.com/news/videos/2026-09-12/ai-agents-test-the-limits-of-human-control-video . FLI's standing "Statement on Superintelligence" (Oct 2025): "We call for a prohibition on the development of superintelligence not lifted before there is: one, broad scientific consensus that it will be done safely and controllably; and two, strong public buy-in." https://futureoflife.org/press-release/prominent-scientists-faith-leaders-policymakers-and-artists-call-for-a-prohibition-on-superintelligence/ . FLI's Summer 2026 AI Safety Index notes every lab "pause" pledge is "heavily conditioned." https://futureoflife.org/wp-content/uploads/2026/07/AI-Safety-Index-Summer-2026-Digital.pdf . A Senate "superintelligence ban / pause" bill from the progressive side is referenced by Semafor (10 Sep) but unnamed. **[UNVERIFIED bill]**
- **12 Sep 2026 — Dario Amodei, "We Must Pace the Frontier"** (the pro-restriction anchor text): "We have always devoted a substantial fraction of our efforts to ... advocating for well-considered regulation of AI, even when this gets us accused of hype, 'doomerism', or regulatory capture." ... "The most effective method of pacing is via regulation that targets all US frontier AI companies." Proposes "permanent embedded evaluators" and a government antitrust waiver for inter-lab safety coordination. https://darioamodei.com/post/we-must-pace-the-frontier ; X: https://x.com/DarioAmodei/status/2098773920774074715 (Sat Sep 12 14:01 UTC). Sam Altman agreed the same day https://x.com/sama/status/2098811563415150910 . Earlier (17 Aug, X exchange with Gavin Baker): "Open-weights do help some with this but are nowhere near a sufficient solution because they simply shift the concentration somewhat to those with the most compute and chips." https://thenewstack.io/amodei-open-weights-compute-regulation/

---

## 5. International picture

### EU
- **Digital Omnibus on AI (Reg. (EU) 2026/1744)**: political agreement 7 May 2026; published OJ 24 Jul; **in force 27 Jul 2026**. It delays high-risk obligations (Annex III → 2 Dec 2027; Annex I → 2 Aug 2028) but **does not touch the GPAI chapter**: GPAI obligations applied 2 Aug 2025; AI Office enforcement powers (incl. fines) began **2 Aug 2026**; legacy models must comply by 2 Aug 2027. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai ; https://ailawdecoded.com/p/eu-ai-act-amended-the-digital-omnibus
- **Open-source exemption (Art. 53(2))**: providers of free-and-open-source GPAI models whose "parameters, including the weights, the information on the model architecture, and the information on model usage, are made publicly available" are exempt from Art. 53(1)(a)-(b) documentation duties — but copyright policy and training-data summary still apply, and "This exception does not apply to general-purpose AI models with systemic risk." Commission Q&A: "open sourcing advanced general-purpose AI models may indeed yield significant societal benefits, including through fostering AI safety research; at the same time, when such models are open-sourced, risk mitigations are more easily circumvented or removed." https://digital-strategy.ec.europa.eu/en/node/13148/printable/pdf ; guidelines https://digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-providers
- **EU reaction to US moves on Chinese models**: no formal Commission statement found. Adjacent signal: FT (9 Sep) reported Anthropic declined to submit its latest model to the UK AISI pre-release, "prompting fears inside the UK government that tech companies are falling into line with the US administration's AI protectionism" (as relayed by Global Times, below). **[UNVERIFIED FT primary]**
- European sovereignty/open-weight actors: Mistral (open-weight, new round Sep 2026), OpenEuroLLM consortium (planned mid-2026 multilingual open models), Cohere–Aleph Alpha (closed, on-prem). https://flintbrief.com/articles/cohere-aleph-alpha-eu-sovereign-ai-2026

### China
- **Releases Aug–Sep 2026 (all open-weight):** Qwen3.8-Max GA 3 Aug (2.4T MoE, 1M ctx; custom licence); DeepSeek V4 Pro GA 13 Aug (1.6T/49B, MIT); GLM-5.3 14 Aug (753B/40B); GLM-5.3-Flash 26 Aug (320B/18B, multimodal); Qwen3.8-Flash 26 Aug (125B/6B, Qwen4 architecture preview); **DeepSeek V4.1 Flash 10 Sep** (552B MoE, 1M ctx, MIT). MiniMax M3 (May/Jun) is the base of Saudi humain-m3. Kimi K3 (16 Jul; weights 26 Jul; 2.8T) remains the largest. https://llm.okamomedia.tokyo/en/china-ai/ ; https://www.intelligentliving.co/deepseek-v4-1-flash-vs-glm-5-3-flash/
- **Benchmarks — open beating closed US models (self-reported unless noted):** DeepSeek V4.1 Flash vs GPT-5.6 Sol / Claude Opus 5: Terminal-Bench 2.1 90.6 vs 88.8/89.1; DeepSWE v1.1 74.2 vs 73.0/74.0; CyberGym 88.1 vs 84.5; AutomationBench 54.8 vs 45.8/50.3; Agents' Last Exam 31.8 vs 26.7/28.6; Codeforces 3471. Closed models still lead GPQA Diamond (94.1 vs 90.9), Terminal-Bench 4.0 (51.8 vs 31.2), HLE no-tools (56.3 vs 36.8). https://www.ashna.ai/blogs/deepseek-v4-1-flash-launch-benchmarks ; https://tradepoint.io/deepseek-ai-released-deepseek-v4-1-flash-with-1m-context-fp4-kv-cache-and-cross-layer-attention-reuse/
- **Independent index (Artificial Analysis, 31 Aug snapshot):** top open-weight models Kimi K3 (60), GLM-5.3 (60), Qwen3.8 2.4T (58), GLM-5.3-Flash (57), DeepSeek V4 Pro (53) — all Chinese; best US open model Nemotron 3 Ultra 38; closed leader Claude Opus 5 63.1. "you can download them." https://ai2.work/blog/chinese-labs-own-the-open-weight-leaderboard-glm-deepseek-qwen
- **Usage:** OpenRouter — Chinese models out-used US models for the 19th consecutive week (31 Aug–6 Sep); four of the top five models Chinese; Tencent Hunyuan Hy4-preview #1 at 14.7T tokens/week (Global Times citing OpenRouter — **[secondary]**).
- **Government / official statements:**
  - **8 Sep 2026** — NSA/FBI/CISA joint advisory "China-Based Artificial Intelligence Companies Conducting Industrial-Scale Distillation Campaigns Against U.S. AI Companies," naming DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, Z.AI; says distillation is the "critical core" of their strategy. https://www.secureworld.io/industry-news/china-distillation-us-frontier-ai-models ; https://www.caixinglobal.com/2026-09-10/in-depth-ai-distillation-in-china-leaves-us-tech-giants-at-odds-102483506.html
  - **9 Sep 2026** — MOFCOM spokesperson: allegations "groundless and lacked legal basis"; distillation "a common practice for models to learn from one another ... essentially a neutral technical approach used by model developers around the world, including US companies"; if the US acts "under the pretext of countering distillation, China will take resolute countermeasures." MOFCOM (July) also said US firms have distilled Chinese models. https://www.globaltimes.cn/page/202609/1370195.shtml
  - China's WAIC (Jul 2026) action plan calls for "broader sharing of open-source AI ecosystems." (Global Times, same.)
- **US–China AI talks:** Reuters (4 Sep) — first bilateral AI-only dialogue planned mid-September, Bessent leading; White House official: "there is currently no planned AI-related meeting in mid-September." Xi visits Washington **24 Sep** with AI "expected to top the agenda." https://www.reuters.com/legal/litigation/us-china-gear-up-mid-september-ai-safety-dialogue-2026-09-04/ ; https://www.latimes.com/politics/story/2026-09-11/... . Bessent (Jul 21): "We are finding watermarks of our U.S. large language models on many of the Chinese models, and that's unacceptable." https://www.cnbc.com/2026/07/21/bessent-china-ai-sanctions.html ; later: "we like open source, but open source has got to be legal – it's not an excuse for IP theft." https://www.foxbusiness.com/economy/bessent-highlights-trump-economy-warns-china-has-done-lot-kicking-lately . **Status as of 13 Sep: talks not confirmed to have happened. [UNVERIFIED]**
- **Trump, 11 Sep**: "It's going to be fine ... We'll always have something to stop them. We'll have a little gear. Boom." and "I have concerns that if we don't win in AI, we're going to be put in a very bad position." (LA Times.) The administration's pre-release framework (June) "applies only to closed models" per CEPA. https://cepa.org/article/ai-confusion-washington-struggles-to-respond-to-china/

### Canada (for the 28 Sep Montreal talk)
- **5 Jun 2026 — national AI strategy launch, Minister Evan Solomon (Montreal)**: "Sovereignty is not a slogan. It's compute, cloud, data, energy, and talent. And if we don't build the capacity here, we're going to depend on the infrastructure, the rules, and the priorities of somewhere else." On open source: "Open source is important option as well ... we got to do that safely and reliably in a Canadian way to reduce cost and increase transparency, support evaluation, give SMEs, nonprofits, researchers, and public interest innovators choice ... We're going to lead a global effort to build open source AI in the public interest. And here in Montreal, that matters." https://www.youtube.com/watch?v=y2sHy83oNuQ . Michael Geist: the strategy "commits to advancing open-source AI as a counterweight to dependence on a handful of proprietary systems." https://mgeist.substack.com/p/ai-for-all-details-to-follow-government
- **Sovereign compute:** $1B public supercomputer (RFP closed 1 Jun; decision "in the fall"); "Enabling Large-Scale Sovereign AI Data Centres" (up to 850 MW → 2.2 GW); AI Compute Access Fund ($300M; 44 projects/$66M, 12 May; 50¢/$ subsidy, 67¢ if Canadian compute); TELUS Kamloops/Vancouver sovereign data centres. https://islandsocialtrends.ca/federal-ai-minister-solomon-on-data-centre-cluster-sovereign-ai/ ; https://www.youtube.com/watch?v=P4japLxun7o
- **Solomon (BetaKit Q&A):** "sovereignty is not solitude" ... "Canada is one of the four countries with a large language model. The US, China, and France have one, and we have Cohere ... Which is why we did an MOU with Cohere, and a sovereign tech alliance with Germany, which led to the Cohere Aleph Alpha merger." https://betakit.com/qa-minister-evan-solomon-on-why-ottawa-wants-canadians-to-adopt-ai/
- **Tension worth naming in Montreal:** Ottawa's strategy is pro-open-source; its national champion (Cohere) is closed-weight; its most famous researcher (Bengio) argues for pacing and safety cases. No Canadian government statement on US restrictions on Chinese models found. **[gap]**

### UAE
- **Sep 2026 — MBZUAI IFM "K2 Horizon"**: six *fully open* models (0.9B–375B; weights, code, data, methodology; Apache 2.0) — "the largest fully open model release in AI history"; explicitly positioned against the "open weights" framing. https://mbzuai.ac.ae/news/mbzuais-institute-of-foundation-models-launches-k2-horizon-the-worlds-largest-fully-open-ai-models-in-history/ ; https://www.wired.me/story/forget-open-weights-the-uae-just-released-six-fully-open-ai-models
- **US–UAE chip deal status:** BIS "Enhanced Favorable Treatment for the UAE" rule effective 10 Jul 2026; G42/Core42 licence-free access "expires on 6 April 2027 unless the companies become US-headquartered." **4 Sep 2026** — Bloomberg: G42 executives held exploratory talks on selling a majority stake to US companies to secure chips. https://www.bloomberg.com/news/articles/2026-09-04/abu-dhabi-s-g42-weighs-us-ownership-to-safeguard-ai-chip-access ; https://thenextweb.com/news/g42-us-ownership-chip-access-export-control-endgame ; https://vision2030.ai/geopolitics/us-chip-export-controls-saudi-uae-divergence/ . Falcon (TII) — no notable Aug–Sep release found. **[gap]**

### UK
- **23 Jul 2026 — UK AISI / US CAISI joint preliminary assessment of Kimi K3 (cyber):** ExploitBench 32% vs ~76% for leading US models (0/41 arbitrary code execution vs 20/41); "The Last Ones" step 17/32 vs 28.5; "Kimi K3's safeguards allow assistance with agentic cyber exploit development." Open-model lag estimated at 4–7 months. Pro-open reading: the best Chinese open model is *behind* frontier closed models on the capability regulators fear most. https://www.aisi.gov.uk/blog/preliminary-assessment-of-kimi-k3s-cyber-capabilities ; https://the-decoder.com/kimi-k3-trails-frontier-us-models-by-a-wide-margin-on-cyber-exploits-and-distillation-may-explain-why/
- Anthropic reportedly declined UK AISI pre-release testing (FT, 9 Sep) **[UNVERIFIED primary]**.

### France, India, Saudi
- **France:** Mistral's Sep 2026 round (NVIDIA, BNP) and Mensch's sovereignty framing (Section 3). Solomon counts France among "four countries with a large language model."
- **India:** Sarvam-30B and Sarvam-105B open-sourced under Apache 2.0, trained "entirely in India on compute provided under the IndiaAI mission," weights on AI Kosh and Hugging Face. https://www.sarvam.ai/blogs/sarvam-30b-105b (release predates window; still the reference point).
- **Saudi (HUMAIN):** **3 Sep 2026, LEAP Riyadh** — humain-m3, a 428B MoE Arabic model "commissioned by HUMAIN and delivered by MiniMax," built on Chinese MiniMax-M3; HUMAIN claims 89.37% on OALL v2 vs GPT-5.6 Sol 87.30% / Claude Opus 5 87.34%; weights "under the MiniMax Community License ... targeted for next month." A US-allied Gulf state building its national model on a Chinese open-weight base is a strong "you cannot pull the ladder up globally" datapoint. https://www.prnewswire.com/news-releases/humain-unveils-humain-m3-a-frontier-arabic-language-model-developed-by-minimax-in-research-preview-on-humain-node-302869158.html ; https://theroboticsmedia.com/article/humain-humain-m3-arabic-frontier-model-minimax-m3-leap-2026-september-3-2026 . HUMAIN's US chip authorisation: 35,000 GB300-equivalents (Nov 2025), still operative.

---

## 6. The pro-open side's core arguments (12), each with a source

1. **Competition / anti-concentration** — "concentrating advanced AI capabilities behind a small number of closed models ... results in a small number of single points of failure, weakens competition." (Letter, 24 Jul.) Sacks: closed labs "already a duopoly in terms of AI model revenue, want the government to eliminate their open-source competition" (19 Jul).
2. **Regulatory capture** — Sacks (17 Aug): identical rules for open and closed "could achieve a similar result" to a ban; "gatekeeping power in a federal bureaucracy working hand-in-glove with a small number of frontier labs." Ng (Ai4): "lobbying ... and the fear-mongering." Chamath (12 Sep): "concentrate enormous technological and economic power with Anthropic."
3. **Security through transparency / defenders need frontier open models** — Letter: "transparency can be more secure than obscurity"; Huang (3 Sep): "defenders need access to frontier capability open models"; Delangue (3 Sep): "we couldn't defend ourselves with proprietary closed source APIs, so we had to use open models."
4. **Balance of power / no benevolent singleton** — Zuckerberg (10 Aug): "There is no such thing as a singular benevolent superintelligence ... The solution is to ensure that superintelligence is broadly distributed." Sacks: "too powerful to centralize."
5. **Sovereignty and lock-in** — Mensch (Sep): "sovereignty, control, choice and independence"; Solomon (Jun): "if we don't build the capacity here, we're going to depend on the infrastructure, the rules, and the priorities of somewhere else"; HF Spring 2026: open weights let governments "fine-tune systems on local data under national legal frameworks."
6. **Science requires more than weights** — Ai2 (Sep): "the evidence you need is upstream of the final checkpoint"; Goodfire/Olmo (9 Sep) traced a safety regression to specific training examples — only possible with open data.
7. **"The weights are already out" / too late** — Hinton (Ai4, Aug): "I think that battle's been lost ... It's too late." Casado (12 Sep): the technology "refuses to be caged."
8. **China will not stop; restriction just moves the frontier abroad** — Letter: restrictions "drive innovation overseas"; Sacks: "island of costly closed models while the rest of the world races ahead"; evidence: Chinese labs hold the top five open-weight slots (AA index) and out-use US models on OpenRouter for 19 weeks; DeepSeek V4.1 Flash (10 Sep) beats Opus 5 / GPT-5.6 on agentic benchmarks.
9. **Soft power / adoption in the Global South** — Ng (Ai4): "AI is a tremendous source of soft power. You can see the way China's model has tremendous accomplishment with Africa"; Saudi HUMAIN building on MiniMax is the live example.
10. **Distillation is legitimate learning, not theft** — Letter: "policymakers should be careful not to conflate legitimate model-development techniques with misappropriation"; Zuckerberg: "protect the principle that you can learn from anything you can observe"; MOFCOM (9 Sep): "a neutral technical approach used by model developers around the world, including US companies."
11. **Cost, diffusion and economic sustainability** — Letter: open weights let organisations "match the right model to the right job at the right cost"; GLM-5.3-Flash / DeepSeek Flash undercut closed APIs by ~10x (intelligentliving, Sep).
12. **The capability gap on dangerous tasks is real and measurable — so target harms, not openness** — UK AISI/CAISI (23 Jul): Kimi K3 far below US frontier on exploit development; Letter: "protections tied to real and demonstrated harms rather than assuming that closed systems are safer by default"; Li (Ai4): "openness doesn't have to be an all-or-nothing choice."

Counter-arguments the talk should anticipate (from the pro-restriction side): irreversibility ("Once released, the weights are beyond the original developer's control" — conceded in the letter itself); removable safeguards (EU Recital 112; Hinton); Amodei's "open weights ... simply shift the concentration somewhat to those with the most compute and chips"; Bengio's demand for a "strong safety case that convinces independent experts" before deployment; Russell's licensing/registration/termination triad.

---

## 7. Unverified, disputed or thinly sourced items

- **Exact date OpenAI signed the letter** — between 25 and 27 Jul 2026 per the explainx tracker; no OpenAI announcement found.
- **X permalinks for Casado, Chamath, Bindu Reddy, Rauch, Delangue (12 Sep)** — quotes taken from Techmeme's aggregation pages; status IDs not retrieved so fxtwitter verification was not possible. Verified via fxtwitter: Sacks 17 Aug (2089227290769080656), Amodei 12 Sep (2098773920774074715), Altman 12 Sep (2098811563415150910).
- **Andreessen / Horowitz / Midha / Boyle Aug–Sep quotes** — none located; a16z's position is inferred from the letter, PAC funding, the a16z paper and Casado.
- **Leading the Future "open source" plank** — not stated in public materials; PAC is pro-preemption/anti-restriction generally.
- **Bengio on open weights in Aug–Sep** — his 11 Sep post is about misalignment/pacing and does not mention open weights.
- **EleutherAI (Biderman) claim that Ai2 disbanded the OLMo team / no OLMo 4** — contradicted by Ai2's Sep statements that the next Olmo generation is in progress. Disputed.
- **US–China mid-September AI talks** — reported by Reuters (4 Sep) via anonymous sources; White House denied a scheduled meeting; not confirmed to have occurred by 13 Sep.
- **Anthropic declining UK AISI pre-release testing (FT, 9 Sep)** — seen only via Global Times summary; primary FT article not fetched.
- **Kimi K3 distilled from Anthropic "Fable" (Kratsios, June)** — a US government allegation; Moonshot denies; UK AISI/CAISI results are consistent with, but do not prove, distillation.
- **DeepSeek V4.1 Flash benchmark wins** — DeepSeek's own launch table; Artificial Analysis independent index puts V4.1 Flash at 40 (below GLM-5.3-Flash 42 and far below Opus 5 63.1). Use "beats on selected agentic/coding benchmarks," not "beats closed models."
- **humain-m3 OALL numbers** — HUMAIN-run, not on the independent Open Arabic LLM Leaderboard; weights not yet released.
- **Senate bill scope re: open weights** — no text; all reporting anonymous.
- **Senate "superintelligence ban / pause" bill** — referenced by Semafor without sponsor/name.
- **Nick Clegg successor (Joel Kaplan) statements** — none found in window.
- **Statements from EFF, OSI, Mozilla, Databricks, Perplexity, Falcon/TII in Aug–Sep 2026** — none found; treat as gaps.

---

## 8. Source list

Primary documents
- Open Weights and American AI Leadership (letter, 24 Jul 2026; V8b roster Aug 2026) — https://www.microsoft.com/en-us/corporate-responsibility/wp-content/uploads/2026/08/open-weight-models-letter_V8b.pdf ; community edition https://openweights.gitlawb.com/letter.pdf
- Anthropic, "Our position on open-weights models" (27 Jul 2026) — https://www.anthropic.com/news/position-open-weights-models
- Dario Amodei, "We Must Pace the Frontier" (12 Sep 2026) — https://darioamodei.com/post/we-must-pace-the-frontier ; X https://x.com/DarioAmodei/status/2098773920774074715
- Sam Altman reply (12 Sep 2026) — https://x.com/sama/status/2098811563415150910
- David Sacks thread (17 Aug 2026) — https://x.com/DavidSacks/status/2089227290769080656 ; (19 Jul 2026) https://x.com/DavidSacks/status/2078826291638522127
- H.R. 10152 — https://www.congress.gov/bill/119th-congress/house-bill/10152 ; text https://www.congress.gov/bill/119th-congress/house-bill/10152/text
- Mark Zuckerberg, "The Future is for Everyone" (10 Aug 2026) — https://about.fb.com/news/2026/08/the-future-is-for-everyone/
- Yoshua Bengio, "Why are AI agents lying, cheating and coordinating?" (11 Sep 2026) — https://yoshuabengio.org/en/blog/why-are-ai-agents-lying-cheating-and-coordinating
- OpenAI / Chris Lehane, "The AI policy window" (9 Sep 2026) — https://www.openai.com/index/ai-policy-window/
- a16z, "Asserting American Leadership in Open Source AI" — https://a16z.com/asserting-american-leadership-in-open-source-ai/
- Andrew Ng, Senate AI Insight Forum written statement — https://aifund.ai/insights/insights-written-statement-of-andrew-ng-before-the-u-s-senate-ai-insight-forum/
- Clem Delangue LinkedIn (19 Jul 2026) — https://www.linkedin.com/posts/clementdelangue_there-have-been-rumors-that-open-source-ai-activity-7484667810909069312-uKfW
- Arthur Mensch LinkedIn (Sep 2026) — https://www.linkedin.com/posts/arthur-mensch_three-years-ago-we-started-mistral-with-activity-7502960387319820288-GscW
- Hugging Face, State of Open Models Summer 2026 — https://huggingface.co/blog/state-of-open-models-summer-2026 ; Spring 2026 — https://huggingface.co/blog/huggingface/state-of-os-hf-spring-2026 ; Solaiman "Open v closed is so 2024" — https://www.linkedin.com/posts/irene-solaiman_open-v-closed-is-so-2024-new-paper-on-considerations-activity-7317942261634232323-EC1M
- Ai2 / Goodfire (9 Sep 2026) — https://allenai.org/blog/goodfire-olmo ; Ai2 Q&A — https://letsdatascience.com/news/ai2-tells-lds-what-open-weights-alone-cannot-answer-198ee205
- OSI Open Source AI — https://opensource.org/ai ; Mozilla Senate statement — https://blog.mozilla.org/en/mozilla/ai/mozilla-joins-latest-ai-insight-forum/
- UK AISI / CAISI Kimi K3 assessment (23 Jul 2026) — https://www.aisi.gov.uk/blog/preliminary-assessment-of-kimi-k3s-cyber-capabilities ; NIST mirror https://www.nist.gov/news-events/news/2026/07/uk-aisi-caisi-preliminary-assessment-kimi-k3s-cyber-capabilities
- EU AI Act / Omnibus — https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai ; GPAI guidelines https://digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-providers ; GPAI Q&A https://digital-strategy.ec.europa.eu/en/node/13148/printable/pdf ; timeline https://ailawdecoded.com/p/eu-ai-act-amended-the-digital-omnibus ; https://docs.modulos.ai/frameworks/eu-ai-act/commission-guidance/gpai-models
- FLI superintelligence statement — https://futureoflife.org/press-release/prominent-scientists-faith-leaders-policymakers-and-artists-call-for-a-prohibition-on-superintelligence/ ; AI Safety Index Summer 2026 — https://futureoflife.org/wp-content/uploads/2026/07/AI-Safety-Index-Summer-2026-Digital.pdf
- MBZUAI K2 Horizon — https://mbzuai.ac.ae/news/mbzuais-institute-of-foundation-models-launches-k2-horizon-the-worlds-largest-fully-open-ai-models-in-history/
- HUMAIN humain-m3 (3 Sep 2026) — https://www.prnewswire.com/news-releases/humain-unveils-humain-m3-a-frontier-arabic-language-model-developed-by-minimax-in-research-preview-on-humain-node-302869158.html
- Sarvam 30B/105B — https://www.sarvam.ai/blogs/sarvam-30b-105b
- Cohere–Aleph Alpha — https://cohere.com/blog/cohere-alephalpha-join-forces
- Evan Solomon, AI strategy launch (5 Jun 2026) — https://www.youtube.com/watch?v=y2sHy83oNuQ ; Compute Access Fund (12 May 2026) — https://www.youtube.com/watch?v=P4japLxun7o

Press and analysis
- CNBC Squawk Box transcript, Huang & Delangue (3 Sep 2026) — https://www.cnbc.com/2026/09/03/cnbc-exclusive-transcript-nvidia-founder-ceo-jensen-huang-and-hugging-face-ceo-clment-delangue-speak-with-cnbcs-becky-quick-on-squawk-box-today.html
- TechCrunch, Ai4 panel Hinton/Li/Ng (12 Aug 2026) — https://techcrunch.com/2026/08/12/as-ai-safety-concerns-mount-three-pioneers-make-the-case-for-staying-open/
- TechCrunch, Amodei pacing (12 Sep 2026) — https://techcrunch.com/2026/09/12/anthropic-ceo-outlines-plan-to-pace-the-frontier/
- Techmeme aggregation (12 Sep 2026) — https://www.techmeme.com/260912/p14 ; https://www.techmeme.com/260912/p13
- The New Stack, Amodei on open weights (17 Aug 2026) — https://thenewstack.io/amodei-open-weights-compute-regulation/
- Reuters, Senate duty of care (11 Sep 2026) — https://www.reuters.com/legal/litigation/us-senate-negotiators-consider-requiring-ai-firms-mitigate-known-major-risks-2026-09-11/
- Reuters, US–China AI talks (4 Sep 2026) — https://www.reuters.com/legal/litigation/us-china-gear-up-mid-september-ai-safety-dialogue-2026-09-04/
- Semafor (10 Sep 2026) — https://www.semafor.com/article/09/10/2026/bipartisan-ai-safety-bill-gains-momentum-on-the-hill
- Nextgov (Sep 2026) — https://www.nextgov.com/artificial-intelligence/2026/09/lawmakers-clash-safety-testing-language-development-ai-legislation-people-familiar-say/415948/
- Progressive Robot tracker (12 Sep 2026) — https://www.progressiverobot.com/2026/09/12/senate-ai-bill-negotiators-ai-firms-mitigate-known-major-risks/
- LA Times, Trump "It's going to be fine" (11 Sep 2026) — https://www.latimes.com/politics/story/2026-09-11/its-going-to-be-fine-trump-rebuffs-rising-alarm-over-ai-dangers-sparking-outcry
- Forbes, Musk vs Coxon (10 Sep 2026) — https://www.forbes.com/sites/siladityaray/2026/09/10/musk-touts-psy-op-and-mocks-ex-anthropic-staffer-who-warned-ai-could-kill-us-all/
- Guardian, Stuart Russell (11 Aug 2026) — https://www.theguardian.com/commentisfree/2026/aug/11/openai-anthropic-google-deepmind-letter
- Bloomberg, Tegmark (12 Sep 2026) — https://www.bloomberg.com/news/videos/2026-09-12/ai-agents-test-the-limits-of-human-control-video
- Bloomberg, G42 US ownership (4 Sep 2026) — https://www.bloomberg.com/news/articles/2026-09-04/abu-dhabi-s-g42-weighs-us-ownership-to-safeguard-ai-chip-access ; TNW — https://thenextweb.com/news/g42-us-ownership-chip-access-export-control-endgame ; Vision2030 chip-rules analysis — https://vision2030.ai/geopolitics/us-chip-export-controls-saudi-uae-divergence/
- Global Times, MOFCOM response (9 Sep 2026) — https://www.globaltimes.cn/page/202609/1370195.shtml
- Caixin, distillation (10 Sep 2026) — https://www.caixinglobal.com/2026-09-10/in-depth-ai-distillation-in-china-leaves-us-tech-giants-at-odds-102483506.html ; SecureWorld on the 8 Sep advisory — https://www.secureworld.io/industry-news/china-distillation-us-frontier-ai-models
- CNBC, Bessent sanctions (21 Jul 2026) — https://www.cnbc.com/2026/07/21/bessent-china-ai-sanctions.html ; Fox Business — https://www.foxbusiness.com/economy/bessent-highlights-trump-economy-warns-china-has-done-lot-kicking-lately
- Axios, administration battle (20 Jul 2026) — https://www.axios.com/2026/07/20/ai-us-china-open-source-kimi ; CEPA — https://cepa.org/article/ai-confusion-washington-struggles-to-respond-to-china/ ; Vorp Labs tracker — https://vorplabs.com/ai-regulatory-updates/open-weights-us-policy ; Turiloop tracker — https://turiloop.com/blog/us-ban-chinese-ai-models
- Politico, Little Tech letter (22 Jul 2026) — https://www.politico.com/news/2026/07/22/startup-founders-urge-trump-not-to-shut-off-chinese-open-weight-ai-01008992
- explainx letter tracker (24 Jul–3 Aug 2026) — https://www.explainx.ai/ (cached: "Open Weights Letter: OpenAI + Google Join — 70+ Signers")
- MIT Technology Review, LeCun/AMI Labs (22 Jan 2026) — https://www.technologyreview.com/2026/01/22/1131661/yann-lecuns-new-venture-ami-labs/
- Business Insider, PAC spending (Jul 2026) — https://www.businessinsider.com/ai-midterm-elections-spending-super-pacs-greg-brockman-anthropic-2026-7 ; Implicator — https://www.implicator.ai/anthropic-donates-20-million-to-counter-openai-backed-super-pacs-in-2026-midterms/ ; Built In — https://builtin.com/artificial-intelligence/ai-super-pacs-2026-midterm-election-regulation
- Chinese model benchmarks — https://www.ashna.ai/blogs/deepseek-v4-1-flash-launch-benchmarks ; https://tradepoint.io/deepseek-ai-released-deepseek-v4-1-flash-with-1m-context-fp4-kv-cache-and-cross-layer-attention-reuse/ ; https://ai2.work/blog/chinese-labs-own-the-open-weight-leaderboard-glm-deepseek-qwen ; https://www.intelligentliving.co/deepseek-v4-1-flash-vs-glm-5-3-flash/ ; https://llm.okamomedia.tokyo/en/china-ai/ ; Kimi K3 — https://felloai.com/kimi-k3/
- The Decoder, Kimi K3 cyber — https://the-decoder.com/kimi-k3-trails-frontier-us-models-by-a-wide-margin-on-cyber-exploits-and-distillation-may-explain-why/
- Canada — https://mgeist.substack.com/p/ai-for-all-details-to-follow-government ; https://betakit.com/qa-minister-evan-solomon-on-why-ottawa-wants-canadians-to-adopt-ai/ ; https://islandsocialtrends.ca/federal-ai-minister-solomon-on-data-centre-cluster-sovereign-ai/ ; The Logic — https://thelogic.co/news/cohere-aleph-alpha-merger-sovereign-ai/ ; BetaKit — https://betakit.com/cohere-to-acquire-germanys-aleph-alpha-in-sovereign-ai-play/ ; Flint Brief — https://flintbrief.com/articles/cohere-aleph-alpha-eu-sovereign-ai-2026
- Narayanan & Kapoor — https://www.normaltech.ai/p/do-ai-risks-require-extraordinary
- Tegmark vs Dean Ball debate — https://lironshapira.substack.com/p/max-tegmark-vs-dean-ball-debate-ban-superintelligence
- HUMAIN analysis — https://theroboticsmedia.com/article/humain-humain-m3-arabic-frontier-model-minimax-m3-leap-2026-september-3-2026 ; WIRED ME on K2 Horizon — https://www.wired.me/story/forget-open-weights-the-uae-just-released-six-fully-open-ai-models
- EleutherAI/Ai2 dispute — https://digg.com/tech/uiog315x